Microsoft Has Spent 280 Billion Dollars on AI. The Guardian Asked Where the Chips Are.

A Guardian investigation published this week, citing internal documents, reports that Microsoft has roughly 2.2 million AI chips installed worldwide after spending around 280 billion dollars since 2022 on land, buildings and computing infrastructure. The point of the story is the gap: experts consulted by the paper estimated that the 5 gigawatts of added data centre capacity Microsoft has described would need something closer to 6.4 million GPUs to fill. Microsoft rejected the arithmetic, calling it “inaccurate, drawing the wrong conclusions from incorrect assumptions.” The company’s shares fell about 3 percent on Monday to close near 480 dollars.

Chief executive Satya Nadella has offered his own explanation of the gap for months, and it is not a chip shortage. It is what he calls warm shells: powered, cooled, finished buildings ready to receive hardware. “It’s not a supply issue of chips,” he said earlier this year, “it’s actually the fact that I don’t have warm shells to plug into.” He has described GPUs sitting in inventory that the company cannot switch on. Delayed sites, including Fairwater in Wisconsin, get pointed at as examples.

Here is why this is more interesting than a corporate spat. For three years the public story of AI scaling has been about chips: who can buy Nvidia’s newest, who is cut off by export controls, who gets to the front of the queue. That framing is quietly going out of date. A modern AI data centre is first an electricity problem, then a construction problem, and only then a semiconductor problem. Substations, grid connections, cooling water and planning permission all move on timescales that measure in years, and none of them speed up because you wired more money. If Nadella’s account is right, the industry’s real bottleneck is now civil engineering and utilities. If the Guardian’s reading is right, then reported capacity figures across the sector may be describing buildings rather than working computers, which makes it harder for anyone outside to judge who actually has what.

A fair caveat: both sides are working from partial information. The Guardian’s estimate of how many GPUs 5 gigawatts implies depends on assumptions about chip generation, power draw per rack and overhead, and reasonable people can land on different numbers. Microsoft has not published its own figure.

What this means for you: very little today, and that is fine. But it is worth updating your mental model. When you read that a company is spending tens of billions on AI, most of that money is not buying chips, it is buying power and property. That is also why AI capacity is arriving unevenly by region, why some services roll out in the US months before Europe, and why questions about data centre electricity use are not going away. If you are choosing cloud providers, “we have capacity” is a claim worth asking follow-up questions about.

Sources

• Inside Telecom: Microsoft AI computing chips anchored by power
• Tom’s Hardware: Microsoft CEO says the company does not have enough electricity to install all the AI GPUs in its inventory
• Data Center Dynamics: Microsoft has AI GPUs sitting in inventory